Choose Power BI Copilot when people should ask about a Power BI report or semantic model inside the Power BI experience. Choose a Fabric Data Agent when the conversation must span one or more supported Fabric data sources and operate as a separately configured data agent. Choose a Copilot Studio custom agent when the job also needs channel delivery, orchestration, non-data knowledge, tools, or business actions. These are not maturity levels. The right choice follows the user’s job, required data boundaries, permissions, and operating model.
All three options still depend on trustworthy definitions and tested source behavior. Moving an ambiguous revenue metric into a more elaborate agent does not resolve the ambiguity.
Decision matrix
| Decision factor | Power BI Copilot | Fabric Data Agent | Copilot Studio custom agent |
|---|---|---|---|
| Primary job | Explore, summarize, or create from Power BI reports and semantic models | Ask questions across supported Fabric data sources through a configured data agent | Deliver a broader conversational workflow that can use knowledge, tools, actions, and other agents |
| Natural home | Power BI | Microsoft Fabric | Copilot Studio and configured channels |
| Data scope | Power BI report and semantic-model context | Supported Fabric sources, including Power BI semantic models, lakehouses, warehouses, SQL databases, Eventhouse/KQL, and selected preview sources | Depends on connected knowledge, tools, and agents; can call a Fabric Data Agent |
| Query path for a semantic model | Copilot generates/uses semantic queries and can render a visual and summary | Data agent invokes its Power BI semantic-model/DAX tooling | Custom agent can delegate to an attached Fabric Data Agent tool, or use other configured tools |
| Best fit | Report consumers or authors already working in Power BI | Data conversations that need a dedicated, governable item across Fabric sources | A business workflow whose analytics answer is one step among routing, knowledge retrieval, and action |
| Main operating concern | Model preparation and capability-specific grounding | Source selection, instructions/examples, publishing/versioning, and source permissions | Orchestration, authentication mode, channel policy, tool permissions, and end-to-end testing |
| Avoid when | Users need a broader cross-system workflow | The use case is already well served inside one Power BI report/model | The only problem is an unprepared semantic model |
This matrix combines current Microsoft product facts with Refinity’s selection guidance. Confirm licensing, capacity, tenant settings, region, preview status, and channel support against current Microsoft documentation before implementation.
Option 1: Power BI Copilot
Power BI Copilot is the most direct choice when the user is already in Power BI and the task is to understand a report, ask a data question, or create/edit report content from a semantic model. Microsoft documents report and semantic-model experiences that can use report visuals, model metadata, generated semantic queries, new visuals, and summaries. Use Copilot with Power BI reports and semantic models
Choose it when
- the governed semantic model is the intended source;
- users work primarily in Power BI;
- answers should preserve the report or model context;
- the main task is analytical exploration or report authoring;
- the model can be prepared and evaluated for the target questions.
Readiness work still required
Microsoft recommends sound model design and the Prep data for AI controls: AI data schema, verified answers, AI instructions, descriptions, testing, and Approved for Copilot. The preparation workflow is currently identified as preview, and output is nondeterministic. Prepare a semantic model for AI
Power BI Copilot is not the right answer merely because the data already appears in a dashboard. Ask whether the target question has an owned metric, explicit calendar and grain, known security behavior, and an expected-answer test.
Option 2: Fabric Data Agent
A Fabric Data Agent is a separately configured Fabric item for conversational access to supported data sources. Microsoft says a data agent can work with sources including Power BI semantic models, lakehouses, warehouses, Fabric SQL databases, KQL/Eventhouse sources, ontologies, and Microsoft Graph in Fabric. A current data agent can combine up to five sources. Add data sources to a Fabric Data Agent
For a Power BI semantic model, the data agent uses a DAX-oriented path and can draw on the model’s measures and business logic. Microsoft’s semantic-model best practices for Data Agent point back to Power BI’s Prep for AI configuration. Semantic-model best practices for Data Agent
Choose it when
- a dedicated conversational data product is preferable to a report-bound experience;
- questions genuinely cross supported Fabric sources;
- the team needs draft/published versions of the agent configuration;
- source selection and agent instructions can be governed as an owned artifact;
- consumers can be given the required permissions on the underlying sources.
Important boundaries
Microsoft says Data Agent executes against the calling user’s accessible data and honors underlying source permissions, including RLS and column-level security for Power BI semantic models. Users need appropriate access to every source a question touches. Fabric Data Agent sharing and permissions
Microsoft also describes current analytical limits: Data Agent retrieves and processes structured data, but does not perform advanced analytics, machine learning, or causal inference. A question such as “What caused the sales spike?” may require evidence and methods outside the agent’s retrieval scope. Create a Fabric Data Agent
Do not select multiple sources simply because the feature permits them. Every additional source adds routing, definition, permission, and evaluation decisions.
Option 3: Copilot Studio custom agent
In this article, “custom agent” means an agent built in Microsoft Copilot Studio—not an unspecified bespoke application. Microsoft describes Copilot Studio as a low-code platform for agents that can use knowledge sources, tools, actions, and channels. A Fabric Data Agent can be attached as a tool so the custom agent can delegate data questions to it. Use a Fabric Data Agent as a Copilot Studio tool
Choose it when
- the experience needs a channel beyond Power BI;
- the conversation must combine analytical answers with policies, documents, or operational context;
- the agent must call tools or perform an action after answering;
- orchestration must decide among several specialized tools or agents;
- the team can govern end-to-end authentication, authorization, and failure handling.
Authentication changes the risk
Microsoft’s current tool integration supports user and maker authentication modes. With user authentication, each user needs access to the Fabric Data Agent and its sources. With maker authentication, users can see results through the maker’s access. That is a material data-boundary decision, not a setup detail. Fabric Data Agent as a Copilot Studio tool
Microsoft also warns that some Fabric Data Agent and Copilot Studio integrations are preview and may have channel or data-boundary considerations. Review the current integration documentation for the intended channel and tenant before release. Connect to a Fabric Data Agent
A decision tree for real use cases
1. Is the job contained within a Power BI report or semantic model?
If yes, begin with Power BI Copilot. Do not add an agent tier until a user or workflow requirement demands one.
2. Must one conversation query several governed Fabric sources?
If yes, evaluate a Fabric Data Agent. First test whether definitions align across those sources and whether target users have the required access.
3. Must the experience combine data with knowledge, routing, channels, or actions?
If yes, evaluate a Copilot Studio custom agent. Use a Fabric Data Agent as a specialized data tool where appropriate instead of recreating semantic query logic inside the orchestrator.
4. Is the main defect actually in the semantic model?
If the model has competing measures, mixed grain, calendar ambiguity, or untested RLS, stop. Repair and test the source before selecting a larger surface.
Worked scenarios
| Scenario | Recommended starting point | Reason | Validation focus |
|---|---|---|---|
| A regional leader asks questions about an existing revenue report | Power BI Copilot | One governed Power BI context and in-product users | Metric/calendar contract, AI schema, verified questions, RLS |
| An analyst asks about warehouse inventory and a Power BI demand model | Fabric Data Agent | The job crosses supported Fabric sources | Source routing, aligned definitions, per-source permissions, cross-source tests |
| A service agent answers a policy question, checks account data, then opens a case | Copilot Studio custom agent with governed tools | Knowledge, data retrieval, orchestration, and an action are all required | Authentication mode, least privilege, action confirmation, audit trail, channel behavior |
| An executive asks “What caused margin to fall?” | No automatic product choice yet | Causal claims require methods and evidence beyond retrieval | Define analytical method, contributing evidence, uncertainty, and human review |
The recommended starting points are Refinity judgments, not Microsoft product guarantees.
Selection questions before architecture
- What exact user decision should the answer support?
- Which source owns the metric definition?
- Is the question single-source or genuinely cross-source?
- Does the user need a visual exploration, a data answer, an action, or all three?
- Which identity should execute each query or action?
- What must the system refuse, clarify, or escalate?
- How will expected answers be tested across roles and prompt variants?
- Who owns the configuration after launch?
If these answers are missing, a platform comparison will create false precision.
Microsoft product facts vs. Refinity recommendations
Microsoft product facts
- Power BI Copilot supports report- and semantic-model-centered experiences.
- Fabric Data Agent supports several Fabric data-source types and can use Power BI semantic models.
- A Fabric Data Agent can be used as a tool by a Copilot Studio agent.
- Permissions and authentication differ by surface and integration.
- Several preparation or integration capabilities are currently in preview or have specific limitations.
Refinity recommendations
- Choose the smallest surface that satisfies the user’s job.
- Treat cross-source scope as a governance cost, not automatically as a benefit.
- Repair semantic ambiguity before changing agent architecture.
- Make identity and data-boundary decisions before building the conversation.
- Evaluate the full path from question to source query to action.
Primary CTA: Use the Power BI Copilot Readiness assessment to separate model problems from architecture decisions.
Related reading: